A Markdown workspace your AI agents can actually use
Connect Claude Code, Cursor, or any MCP client to mdedit.ai. Agents draft and suggest; you approve what lands.
By Vivek Maskara · Last updated
mdedit auth login
claude mcp add --transport stdio mdedit -- npx -y @mdedit/mcp-server
What you'd ask for
Four prompts you can paste into a connected client, and what each one does to your document.
Review this draft and leave suggestions instead of editing directly.
Summarize the open review threads and resolve the ones I already replied to.
Create an article from these meeting notes in my Docs workspace.
Publish this doc as a live link.
The agent proposes, you approve
Suggestions wait for a person
An agent proposal stays anchored to its passage until you accept or reject it in the editor.
Publishing asks first
A public link needs the publishing:write scope and an explicit confirmation on the tool call.
Nothing beyond the grant
Every action is limited to what the key or the sign-in allows, and anything else fails closed.
Connect your client
Two ways in, both running the same fourteen tools. Node.js 20 or newer is required for the local package.
Already use the mdedit.ai CLI? No key needed.
The MCP server reuses your CLI sign-in from the operating system keychain, so nothing secret goes into your client config.
mdedit auth login
claude mcp add --transport stdio mdedit -- npx -y @mdedit/mcp-server
Or create a scoped API key
Works with every MCP client. Create the key in Settings, Integrations, API Keys.
claude mcp add --transport stdio \ --env MDEDIT_API_KEY=mdh_your_key \ mdedit -- npx -y @mdedit/mcp-server
Hosted endpoint for HTTP clients
POST https://mcp.mdedit.ai/mcp
X-API-Key: mdh_your_keyClients that browse the official MCP registry can find the server as io.github.mangoappstudio/mdedit.
Client by client setup lives in the Claude Code guide, and connection troubleshooting in MCP authentication.
Tools at a glance
Read and explore
Find the right workspace and document before touching anything.
list_workspaceslist_articlesread_article
Write, review and suggest
Draft, propose, and work through feedback where your team already reads it.
create_articleedit_articleadd_commentadd_suggestionreply_to_threadresolve_threadlist_review_threads
Publish and presence
Public links stay behind a confirmation, and agents appear alongside your teammates.
publish_articleget_publish_statusunpublish_articleget_presence
Fourteen tools total. Full reference with parameters in the agent docs.
One key per agent, scoped to one job
Keys start with mdh_ and carry a fixed set of permissions.
Review only
articles:read, reviews:writeReads documents and leaves comments and suggestions. It cannot rewrite your prose.
Editor
articles:writeCreates documents and applies anchored edits. A write scope also covers the matching read.
Publisher
publishing:writeCreates or disables a public mded.it link, and only with the confirmation flag set.
A tool the key cannot reach returns a permission challenge instead of doing half the work. Read the API key guide.
Other ways to automate mdedit.ai
Command line
Script workspaces, documents, and exports from your terminal, with JSON output for pipelines.
See the CLIAgent skills
Portable instructions that tell an agent when to save, find, revise, or publish a document. MCP stays the transport underneath.
Browse skillsSetup guides
Claude Code setup, hosted authentication, scopes, and API key rotation.
Read the setup guideMCP server FAQs
The Model Context Protocol is an open standard for connecting AI clients to external tools. An MCP server publishes a list of typed tools, and the client decides which ones to call. @mdedit/mcp-server is the mdedit.ai implementation, and it publishes fourteen tools for Markdown documents.
Yes, if you use the mdedit.ai CLI. Run mdedit auth login once, and the local MCP server picks up that sign-in from your operating system keychain, so no key sits in your client config. MCP clients cannot start a browser sign-in on their own yet, so an API key stays the path that works everywhere else.
Anything that speaks MCP over standard input and output can run the local package, either with the shared CLI sign-in or with MDEDIT_API_KEY in its environment. Claude Code has a documented setup command. Clients that support Streamable HTTP and custom headers can use the hosted endpoint with an API key instead.
Run the local package when your client launches processes, which is the usual case for terminal agents. Use https://mcp.mdedit.ai/mcp with an X-API-Key header when your client only speaks HTTP. Both paths expose the same fourteen tools. See MCP authentication.
No. Publishing needs the publishing:write scope, and publish_article also needs an explicit confirmation flag on the call. An ordinary editing key cannot make a document public. Once a link exists, anyone who has it can read the document, so treat that step as a deliberate one.
The call fails closed. The tool returns a permission challenge naming the missing scope instead of partially completing the work, so a review-only key stays review-only.
No. add_suggestion records a proposed replacement against a passage, and it stays a proposal until a person accepts it in the editor. add_comment never changes prose at all. Only edit_article writes to the document, and it needs a grant scoped for writing.
They serve different jobs, and they share one sign-in. MCP is for an agent working inside a document during a session. The mdedit.ai CLI is for scripts, pipelines, and terminal work where you want plain commands and JSON output.
Give your agents a document to work in
Connect your client in two commands, and keep the final say on every change.